In addition to all that’s been said here already, I’d like to note that models can be used in a variety of ways. For example, you could use AIXI (or AIT concepts more broadly) as a baseline to ask when, and to what extent, RL-trained agents are or are not like AIXI. Certain research programs benefit from assuming that agents are AIXI, but we should not limit ourselves.
In addition to all that’s been said here already, I’d like to note that models can be used in a variety of ways. For example, you could use AIXI (or AIT concepts more broadly) as a baseline to ask when, and to what extent, RL-trained agents are or are not like AIXI. Certain research programs benefit from assuming that agents are AIXI, but we should not limit ourselves.